Energy optimal 3D flight path planning for unmanned aerial vehicle in urban environments

Abstract

This paper presents a general approach to compute energy optimal flight paths for unmanned aerial vehicle (UAV) in urban environments. To minimize the energy required, the flight path is optimized by exploiting local wind phenomena, i.e., upwind and tailwind areas from the airflow around buildings. A realistic wind field of a model urban environment typical for continental Europe is generated using PALM, a Large Eddy Simulation tool. The calculated wind field feeds into the flight path planning algorithm to minimize the energy required. A specifically tailored A-Star-Algorithm is used to optimize flight trajectories. The approach is demonstrated on a delivery UAV benchmark scenario. Energy optimal flight paths are compared to shortest way trajectories for 12 different scenarios. It is shown that energy can be saved significantly while flying in a city using knowledge of the current wind field.

1 Introduction

There are currently big challenges in last-mile logistics in urban environments. Novel approaches are becoming increasingly essential due to rising traffic and space use. The last step of a supply chain is the least efficient. However, electrically powered unmanned aerial vehicles (UAV) are an environmentally and time-efficient alternative in the last step of a supply chain, even in remote locations.

The adoption of last-mile logistics in urban environment by delivery UAVs motivates this work. UAVs can reduce urban street traffic and shorten delivery times. Furthermore, they also have the possibility of automation and are environmentally friendly due to electric powering. In contrast to ground-based vehicles, UAVs have lower payload [1]. Therefore, the efficiency of logistic UAVs has to be increased. An easy approach is optimizing the flight path for an energy optimal flight trajectory. Our previous work in [2] shows that exploiting local wind conditions in an urban environment can significantly reduce the power consumption.

This paper contributes an approach to obtain optimal energy efficient flight paths of typical delivery missions using the knowledge of the wind field in an urban environment. This novel holistic approach consists of different aspects. First, a qualitative realistic wind field for an urban environment is necessary. Various methods ranging from simple statistical methods to complex approximative solutions of the atmospheric boundary layer equations exist in literature. For a good overview, the reader is referred to [3]. In this paper, a Parallelized Large-Eddy Simulation Model (PALM) [4] is used to prognose the wind field, see Sect. 3. It is a Computational Fluid Dynamics (CFD) simulation, or to be more precise a Large Eddy Simulation (LES).

Second, ensuring an energy-optimal flight through different wind conditions is a classical route optimization problem. There are various methods to find an optimal route: Branch-and-Bound [5], evolutionary computing, e.g., [6], multiple-agent-systems [7] neural network [8], and experience optimization can all be used in the embedded optimization process. Common problems in UAV flight path planning and optimization in literature deal with finding the shortest trajectory in scenarios with obstacles [9], hazardous weather avoidance [10], restricted air space [11], agricultural applications, such as fertilizers and pesticide spray, in crop fields just on specific regions [12], and military objectives like avoiding enemy radar sites [13]. A hybrid routing and scheduling problem of an UAV delivery system leads to minimizing the travel time [14]. In [15], modern airline trajectories have been optimized with respect to minimum fuel flow, considering flight paths constraint in the horizontal plane. Most of the trajectory optimizations, e.g., [15,16,17], uses the A-star-algorithm or its variant Theta-Star, e.g., [10, 18]. This type of Branch-and-Bound method has advantages in three dimensional path finding [15] and computational efficiency [19]. A custom-tailored A-star algorithm is proposed in Sect. 4 to deal with the energy efficient path planning in urban environments. It specifically considers turning constraints of the UAV directly in the optimization.

Finally, pruning and smoothing techniques are required to achieve a flyable trajectory. The optimization itself is based on discrete grid points leading to a non-smooth, fractioned track. Pruning is a basic pre-smoothing technique to find shortcuts in grids for unrestricted air space. Path smoothing is defined using piecewise polynomials to achieve a continuous trajectory respecting the limits of the UAV flight envelope. We use continuous cubic Bézier spiral segments, derived in [20] to satisfy maximum curvature constraints of the UAV, see Sect. 4.3.

We generate a realistic city district to showcase the proposed approach applied to a typical European area. Moreover, the applied holistic approach connects the generated wind field, using the Large Eddy Simulation tool PALM, path optimization, and flight trajectory modeling. The methodology is verified by 12 delivery tasks, where their energy optimzied paths are compared to the shortest way in Sect. 5.

2 Problem formulation

A typical last-mile logistic scenario for a UAV is the delivery of goods in a city district flying from a pick-up point to a drop-off point. The route is operated by a fully electric fixed wing aircraft, similar in size and characteristics to a Phoenix Wings PWOne delivery UAV [21] with a wing span of ${1.3}\,\hbox {m}$ and a maximum take-off mass of ${3.4}\,\hbox {kg}$. Average cruise speed is ${60}\,\hbox {kph}$, where an estimated best glide ratio of 20 is assumed.

We develop a generic city model, which represents a typical European urban area. It is consisting of an arrangement of eight uniquely shaped buildings with different building heights, as illustrated in Fig. 1. In detail, there are three residence buildings with a height of ${50}\,\hbox {m}$, four-terraced houses with a height of ${20}\,\hbox {m}$ and a supermarket building with attached office block with a height of ${15}\,\hbox {m}$. This arrangement is developed to obtain a scenario with typical local wind effects. Exploiting these wind fields presents an opportunity to reduce the energy consumption of an UAV during a delivery mission.

Fig. 1

Generic city model, which represent a typical European urban area for simulating delivery tasks

This paper examines four delivery tasks with three different wind speeds each. This results in a total of 12 test scenarios. Figure 2 depicts these scenarios. Only one wind direction is used for all cases to stay in the influence zone of the high-rising buildings. These four delivery tasks are defined by flying from Point North to Point South and South to North, as well as Point West to Point East and East to West. In each scenario, the goal is to minimize the UAV’s energy required using local flow effects to its advantage. It is assumed that the aircraft always flies at a constant true airspeed corresponding to its best-performance cruise speed. For each scenario, the same start and end altitude of ${20}\,\hbox {m}$ are assumed. This represents a realistic initial altitude for take-off and landing in multicopter-mode for air delivery in cities. Take-off and landing procedures are neglected for energy consideration. Average wind speeds in a typical continental European city are the basis for this investigation. Dresden, Germany, represents such a city, and its wind speeds averaged over a whole year are depicted in Fig. 3. At the height of ${10}\,\hbox {m}$, three characteristic wind speeds can be extracted to represent the wind speed specter. The average wind speed of the windiest and calmest days and the average wind speed over one year are summed up in Table 1. They were transformed into a wind profile shape, obtained by a wind tunnel experiment, taken from [22]. The upper end of such a profile is characterized by the undisturbed flow above with its speed. Table 1 summarizes this freestream wind speed $u_{W\infty }$ for the profiles.

Fig. 2

Flight scenario with constant wind direction from west, as well as four tracks by flying from Point West to East, South to North and vice versa for each

Fig. 3

Average of mean hourly wind speeds (dark gray line), with 25th to 75th and 10th to 90th percentile bands for Dresden, representing a typical continental European city [23]

Table 1 Summary of wind speeds applied in scenarios

3 Wind field prediction using LES

The PALM system is used in the present paper to solve the large eddy equations and to obtain a realistic wind field within an urban environment for the following flight path optimization. PALM¹ was developed by the University of Hannover as a tool to simulate urban climate and its boundary layer [4]. It has been validated for flow around solid obstacles [24, 25] and used in various real urban environments [24,25,26,27]. A detailed description of PALM and its capabilities can be found in [4]. In short, PALM calculates the non-hydrostatic, filtered, incompressible Navier–Stokes equations in Boussinesq-approximated form. Furthermore, the subgrid-scale turbulent kinetic energy (SGS-TKE) is solved.

For the presented study, the topography of the generic city model was created in PALM. PALM uses equidistant horizontal grid spacings. Each grid volume can be either air or solid, i.e., representing a building. Due to the rectangular grid, oblique building walls require additional care in modeling. If more than half of a grid volume is geometrically belonging to a building, the whole volume will be set as a solid. An obstacle-free extended computational domain in the vertical, lateral, and flow directions is necessary to catch all flow effects. A guideline for CFD simulation of flows in the urban environment in [28] includes recommendations for the distances from built-up area to boundary. The simulation inlet and the top boundary should be at least five times the height of highest building away from the closest building to comply with this guideline. For urban areas with multiple buildings, the lateral boundaries can be closer and are placed four times the height of highest building away. The region behind the considered area is important to allow redevelopment of the flow behind the wake region. It depends on the blocked cross-sectional area of the buildings in flow direction [28]. Sensitivity studies with different computation domain sizes were conducted to eliminate artificial accelerations to fulfill these requirements. These yield a distance of 13.5 times the height of highest building from last building to outflow. The grid size for all experiments was set to $\Delta _{x,y,z}={2.5}\,\hbox {m}$. This size offered the best trade-off between computational time and accurate resolution of flow effects. The same setup was used for the different wind speeds described in Sect. 2. The setup of the topography for the scenarios is depicted in Fig. 4.

Fig. 4

Scenario for LES simulation with up- and downstream area

For the simulation, a realistic wind profile is very important to achieve meaningful results. PALM is using a cycling boundary condition that uses the outflow as inflow, i.e., the outflow is turbulent after flowing through the urban area and is reused as realistic turbulent inflow until the whole simulation converges. Hence, it is necessary to set an initial wind profile to initialize the simulation. Therefore, a realistic baseline from a scaled wind tunnel experiment for urban environment applications was used. Since the simulation converges to some quasi-steady solution, it has to be checked finally, if the actual wind profile from PALM is still matching with the wind tunnel experiment. For the simulation, the settings shown in Table 2 are used for PALM. The wind tunnel experiment was conducted at Technische Universität Dresden and its data is taken from [22]. Hence, the inflow averaged wind profiles were compared. Figure 5 shows both, rescaled wind tunnel experiment and LES simulation, normalized to $u_{ref}$, the velocity u in the height of $z={30}\,\hbox {m}$. They behave in a good consistency in the range of the building height. The variations above ${50}\,\hbox {m}$ can be neglected as these are outside the considered altitude range.

Table 2 Used settings in the LES model PALM

Fig. 5

Comparison between wind profile of wind tunnel experiment and LES simulation

Before the data is used for trajectory optimization, it has to be ensured that the LES is fully developed. Therefore, it has to be checked, if a quasi-stationary state is reached and the grid spacing is small enough to resolve the turbulent transport with it. Both criteria are satisfied. The time series of kinetic energy E, turbulent kinetic energy $E^*$ and the maximum velocity components show no trends anymore after ${15100}\,\hbox {s}$ ($\approx$ 4 h) of simulation time, so the simulation was converging. Furthermore, the subgrid-scale momentum flux $w''u''$ is one order of magnitude smaller than the resolved-scale counterpart $w^*u^*$ after two hours of calculation time, see Fig. 6. From this, it can be concluded that the grid spacing is sufficiently small according to [29].

Fig. 6

Subgrid-scale momentum flux $w''u''$ and resolved-scale counterpart $w^*u^*$, for each mean value over the height

The path planning algorithm needs the velocity components of the wind at each grid point as input, but PALM provides the perpendicular component at each side face of the volume mesh [30],e.g., component u in Fig. 7. Hence, they are converted. In the converting process, the mean value of one velocity component is generated by respective component of the four adjacent side faces that belong to the point and are perpendicular to the component.

Fig. 7

Converting wind component u from grid data $u_{\text {i*,j*}}$ to point data $u_{\text {i,j}}$

4 Trajectory optimization

The objective of the flight path planning is to minimize a cost function J,

$$\begin{aligned} J = \displaystyle \int _{s_{0}}^{s_{f}}{A(s)} \, \text {ds}, \end{aligned}$$

(1)

where s is the position vector of the aircraft, $s_0$ the starting point, $s_f$ the arrival point and A(s) the quantity to be minimized at each step. To minimize the distance, for instance, A(s) is set to $A(s)=1$. In general, this is a complex optimization problem that is usually solved over a discrete grid instead of the continuous integral. Multiple solvers exist in literature to minimize Eq. (1), such as the Branch-and-Bound algorithm [5] in the present paper or evolutionary computing-based ones [6].

4.1 Cost function

The optimization problem in the present paper is to minimize the energy supplied by the propulsion system for the UAV for a flight with variable altitude between a fixed starting and ending point (see Sect. 2 for the detailed problem). In order to achieve an optimal flight trajectory, a new cost function detailed in this section is proposed. First, the generic flight path cost function (2) is defined as an integral over the flight paths and then discretized in sequentially flown through grid points until ending point N.

$$\begin{aligned} \begin{aligned} J&= \displaystyle \int _{s_{0}}^{s_{e}}{A(s)} \, \text {ds} = \sum _{i=0}^{N} {A_i} \, \Delta s_i. \end{aligned} \end{aligned}$$

(2)

The second step is to derive a function for energy required to substitute the $A_i$ in Eq. (2). First, we consider straightforward a 2D level flight and extend the approach to 3D space afterward. All flight mechanical assumptions are derived from corresponding standard literature, e.g., [31]. For simplification, the energy supplied by the propulsion system is assumed to be proportional to the thrust force multiplied by the path covered by the UAV. For the UAV in steady cruise flight, it is assumed that the thrust force equals the aerodynamic drag force D. Using this assumption, the energy required between two points of the discretized flight path $E_i$ is given by

$$\begin{aligned} \begin{aligned} E_i&= D \cdot \Delta s_i = D \cdot u_ UAV,TAS \cdot \Delta t_i \end{aligned} \end{aligned}$$

(3)

where $u_ UAV,TAS$ is the true airspeed (TAS) of the UAV and $\Delta t_i$ is the time required to cover the distance between two grid points $\Delta s_i$. By introducing $\Delta t_i$, it is possible to implement the wind influence as follows. Defining $u_ UAV,i ^*$ as the ground speed at the i-th grid point, the time it takes to cover $\Delta s_i$ is given by

$$\begin{aligned} \Delta t_i = \frac{\Delta s_i}{u_ UAV,i ^*} \,. \end{aligned}$$

(4)

Considering now headwind $u_{W,i}$ and crosswind $v_{W,i}$ at the i-th grid point, the relationship between ground speed $u_ UAV ^*$ and true airspeed is given by

$$\begin{aligned} u_ UAV,i ^*= \sqrt{(u_{UAV,TAS}-u_{W,i})^2-v_{W,i}^2} \,. \end{aligned}$$

(5)

Using Eqs. (3),  (4) and  (5) yields the following equation for the 2D energy required between two nodes:

$$\begin{aligned} E_i = D \cdot u_ UAV,TAS \cdot \frac{\Delta s_i}{\sqrt{(u_ UAV,TAS -u_{W,i})^2-v_{W,i}^2}} \,. \end{aligned}$$

(6)

In the next step, up- and downwinds are implemented for level flight. E.g., assuming no change in flight height due to upwind components requires the UAV to pitch down to hold the flight level. Hence, the upwind $w_{W,i}$ leads to a decrease of the flight path angle $\gamma _{i,wind}$:

$$\begin{aligned} \sin (\gamma _{i,wind}) = \frac{w_{W,i}}{u_{UAV,i} ^*} \,. \end{aligned}$$

(7)

A decrease of flight path angle leads to less required thrust due to the balances of flight mechanic forces, because a component of the weight supports the force in flight direction. Hence, at an upwind point, the perceived drag that needs to compensated by thrust reduces to:

$$\begin{aligned} D_i = D_0 - \sin (\gamma _{i,wind}) \cdot mg \end{aligned}$$

(8)

with $D_0$ as drag force in normal horizontal flight condition and mg the weight.

Considering now the 3D case allowing changes in altitude, an additional summand is added to the flight path angle $\gamma _{i}$, i.e., $\gamma _i = \gamma _{i,path} + \gamma _{i,wind}$. The wind component is as defined in Eq. (7). The altitude changing component is given by

$$\begin{aligned} \tan (\gamma _{i,path}) = \frac{\Delta s_{i,z}}{\Delta s_{i,xy}} \,, \end{aligned}$$

(9)

where $\Delta s_{i,z}$ is the vertical distance and $\Delta s_{i,xy}$ the horizontal distance to the next point. Implementing this into Eq. (6) and using the relationship in Eq. (8) yields:

$$\begin{aligned}{} & {} E_i = \frac{\Delta s_i}{\sqrt{(u_ UAV,TAS -u_{W,i})^2-v_{W,i}^2}} \cdot \left( D_0-\sin (\gamma _i) \cdot mg \right) \cdot \nonumber \\{} & {} \quad u_ UAV,TAS \,. \end{aligned}$$

(10)

The influence of vertical wind on ground speed $u_ UAV,i ^*$ can be neglected because it is significantly smaller than the other components. The definition of $D_0$ is set by the glide ratio G:

$$\begin{aligned} D_0 = \frac{L}{G} = \frac{mg}{G} \end{aligned}$$

(11)

with L as lift force equal to the weight force and the glide ratio G of the fixed-wing UAV. Using Eq. (10), the cost function for the minimization of the total energy required can be written as:

$$\begin{aligned}{} & {} E = \sum _{i=0}^{N} \frac{\Delta s_i}{\sqrt{(u_ UAV,TAS -u_{W,i})^2-v_{W,i}^2}} \cdot \left( D_0-\sin (\gamma _i) \cdot mg \right) \cdot \nonumber \\{} & {} \quad u_ UAV,TAS \,. \end{aligned}$$

(12)

4.2 Extended A-star-algorithm

As mentioned before, the A-Star-Algorithm is often used to optimize routes or trajectories. First, we want to give a brief explanation of the basic A-Star-Algorithm. Afterward, a custom tailored version is proposed.The basic A-Star-Algorithm is an adaption of the Dijsktra-algorithm [19] and finds cost optimal paths from a starting to an ending point. Alg. A in the appendix shows and explains the exact steps of basic A-Star-Algorithm. In brief, the algorithm starts at the starting point and creates the optimal path by following the best point at a branching. The best point is characterized by the lowest total cost of the path. This cost is defined by summing up the exact cost of the path from the starting point to the node s, g(s), and the heuristic estimated cost from node s to the ending point described by h(s). Hence, the algorithm starts to check all possible paths from the starting point in detail, but stops moving forward with a potential path if the considered branch is considered too expensive. This approach leads to computational savings time because not all possible paths need to be computed.

In this paper, the exact cost function g(s) of the algorithm is set to Eq. (12). The heuristic cost function h(s) provides A-Star an estimation of the minimum cost from any vertex s to the goal. On the one hand, the closer h(s) is to the exact cost of this path, the faster A-Star finds the best way [32]. On the other hand, if h(s) is greater than the exact cost, it is not guaranteed to find the best way. Hence, the heuristic function must be smaller than the exact cost. Heuristic costs must be the same scale as the exact cost. Therefore, in this paper, h(s) is defined on the basis of Eq. (3) as

$$\begin{aligned} h(s) = D_0 \cdot \Delta s_{s,s_{goal}} \cdot w \,, \end{aligned}$$

(13)

where D is the aerodynamic drag force, $\Delta s_{s,s_{goal}}$ the distance from point s to the ending point and w a weighting factor. The weighting factor is set to 0.001 to ensure staying below the exact cost. Equation (13) represents the energy required from point s to the final point if the UAV flies directly without the influence of wind. However, it is possible that this energy required is approximately zero if favorable wind conditions are present. An extensive parameter study showed that $w = 0.001$ always leads to the optimum path under acceptable computation time (in this scenario).

An extension of the basic algorithm enables to consider a maximum curvature constraint of the UAV (specified in Sect. 4.3). At constant speed, an UAV would not be able to turn at smaller radii, e.g., due to a maximum allowed load factor. That imposes the demand for the curvature constraint. This constraint leads to a maximum allowed change of heading at a specific distance. Therefore, at each point s, the information about the point before has to be available to compute the angle between the two connecting lines. This angle is equivalent to the heading change of the UAV. As a consequence, each point s has different outgoing branches to leave s, depending on the direction of reaching this point. Thus, we introduce an extended A-Star-Algorithm. Alg. B in the appendix shows and explains the differences to the basic one, where a function ensures the prespecified requirements of the connecting lines between the points. Furthermore, the basic variable of just one point-to-point connection becomes one that consists of the point connections of three points to consist of the allowed path segment.

4.3 Path modeling

4.3.1 Path discretization

As a starting point in this paper, the vertices from PALM in Sect. 3 are used, namely ${2.5}\,\hbox {m}$ equidistant grid points. In the path finding problem, an aircraft can fly from the current vertex to all adjacent vertices as illustrated in Fig. 8. This figure shows the 3D grid with the current vertex in black, the vertices on the same level in yellow, below them in green, and above them in red. However, an equidistant grid, as used for the wind field calculation, disregards any limits imposed by the UAV’s performance. For instance, an equidistant grid would lead to a flight path angle demand of 45 degrees which well exceeds the UAVs maximum achievable climb speed. In addition to the flight path angle range, a minimum turn radius is considered in this paper. Both lead to a shrinking of the original grid in the vertical direction and an expansion in the horizontal one.

Fig. 8

The connection lines from a vertex (black) to all adjacent vertices (red, green and orange) define possible tracks (blue arrows) in 3D space

Specifically, the grid is adjusted in the following ways. The minimum turning radius $r_{min}$ of the UAV, which is used for the horizontal grid distance, is defined by

$$\begin{aligned} r_{min} = \frac{u_ UAV,TAS ^2}{g \cdot \tan (\arccos (\frac{1}{n_{max}}))} \,, \end{aligned}$$

(14)

where $u_ UAV,TAS$ is the true airspeed (TAS), g the gravitational acceleration, and $n_{max}$ a maximum load factor of the UAV. Reference [33] shows that static load factors of UAVs do not exceed existing aviation regulations. Thus, a load factor of $n=2.5$ is chosen for turning flights. The minimum horizontal flight grid distance $\Delta f_h$ is set to

$$\begin{aligned} \Delta f_h = \frac{r_{min}}{\Delta x} \,, \end{aligned}$$

(15)

where $\Delta x$ can be substituted by $\Delta y$ due to the equidistant grid of the LES.

The optimal gliding ratio G determines the vertical flight grid distance $\Delta f_v$ as:

$$\begin{aligned} \Delta f_v = \frac{\Delta x}{G} \,. \end{aligned}$$

(16)

Eq. (16) assumes flying with the maximum glide ratio as the best flight condition for descent, even if the glide path differs because of the wind influence. Hence, the value of climb ratio is the same as the maximum glide ratio, where it is assumed that the electric propulsion is powerful enough and close to its optimum operation condition for the climbing flight. Given the specifications of the UAV $u_{UAV,\text {TAS}}={60}\,\hbox {kph}$ and $G=20$, as well as LES grid $\Delta x={2.5}\,\hbox {m}$, $\Delta f_h$ is $\Delta f_h = \frac{r_{min}}{\Delta x} = \frac{12.4 m}{2.5 m} \approx 5$ and $\Delta f_v$ is $\Delta f_v = \frac{\Delta x}{\Upsilon } = \frac{2.5 m}{20} = {0.125}\,\hbox {m}$. Since this derived vertical grid is finer than the one, used in PALM, wind components $u_{w}, v_{w}, w_{w}$ are derived by linear interpolation. The horizontal grid is coarser than the LES grid. Nevertheless, the wind field data of the points between it is taken into account for energy required determination.

4.3.2 Path pruning

The resulting path of an A-Star-Optimization run is characterized by successive waypoints. The adjusted grid size as explained in the previous paragraph constrains each step size. However, there can be paths with lower cost by ensuring a direct connection between non-adjacent points. Pruning or sometimes misleadingly called post-smoothing² is a method to find these shortcut after the first optimization constrained to the grid. A basic pruning technique is looping through the path obtained by A-Star-Algorithm and check if there is a straight connection between points without crashing an obstacle (see Alg. C in the appendix). This method has a shortcoming, e.g., if searching for the shortest path. On the one hand, these line-of-sights lead to shorter tracks. On the other hand this algorithm is not able to find the shortest way, if there is a path with more than two convex courses. There will always be a stop at the protruding point, which is illustrated in Fig. 9.

Fig. 9

The basic pruning algorithm is not able to find a direct line if there is a doubled-convex path

The approach in this study is using the A-Star-Algorithm once again for the pruning step. Only the points $s*$ obtained by the first run are considered as vertices for the second run. Each vertex has a connection to every other vertex if there is a line of sight between them. The other A-Star-Inputs are handled as follows. The exact cost $g(s*)$ of going unbounded to the is defined by calculate this direct way. The heuristic cost $h(s*)$ is set by straight-line cost without respecting obstacles like in the first run. The algorithm finds a path that is the true low-cost one. Due to the very low number of considered vertices, this step is computationally fast.

4.3.3 Path smoothing

Finally, the optimal waypoint-based paths have to be converted to a flyable smooth trajectory. Therefore, a continuous-curvature path-smoothing algorithm based on cubic Bezier curves, including a maximum curvature constraint [20], is used. This method is applied to the path in sections. Every section takes three path points into account. Every line between two path points is split into two halves to retrieve turning points of one section with a continuous junction to the next section. On the basis of path point $t_3$, the procedure is illustrated in Fig. 10 and can be described as follows. Consider the section containing the points $t_2$, $t_3$ and $t_4$. The proposed smoothing adds turning points $tp_{23}$ and $tp_{34}$ halfway between $t_2$ and $t_3$ and $t_3$ and $t_4$, respectively. The original point $t_3$ becomes a curvature steering point and hence, is no longer a part of the flight trajectory. The smoothing procedure is detailed in [20]. In short, the points are transformed into one plane to become a 2D problem, first. Second, eight control points are determined that construct two cubic Bézier spiral curves between $tp_{23}$ and $tp_{34}$.

Fig. 10

Path smoothing of gray path between point $t_2$ and $t_4$: Define turning points $tp_{2,3}$, $tp_{3,4}$ and construct two cubic Bézier spiral curves

One curve P is defined by the four control points $B_0, B_1, B_2, B_3$ via the function

$$\begin{aligned}{} & {} P(m) = B_0 (1-m)^3 + 3B_1 (1-m)^2 m \nonumber \\{} & {} \quad \quad \,\, + 3B_2 (1-m) m^2 +B_3 m^3 \,,\, m \in [0, 1] \,. \end{aligned}$$

(17)

Third, the control points are getting transformed back into 3D space. Step 2 contains the consideration of the minimum turning radius constraint. The minimum turning radius $r_{min}$ of the UAV in Eq. (14) provides a maximum curvature $k_{max}$ of the flight path:

$$\begin{aligned} k_{max} = \frac{1}{r_{min}} \end{aligned}$$

(18)

and is considered by

$$\begin{aligned} d_{min} = \frac{1.228 \cdot \sin \beta }{k_{max} \cdot \cos ^2 \beta } \,, \end{aligned}$$

(19)

where $\beta$ is the half angle between the two lines. The distance $d_{min}$ is compared with the length of the two lines $\overline{tp_{2,3} t_3}$ and $\overline{t_3 tp_{3,4}}$. Note that, it is possible that one line is shorter than $d_{min}$. To fulfill the constraint, we introduce several approaches in this order: First, it is checked if skipping the path point, the point before or the following point, will improve the smoothing procedure. Second, it is checked if slightly shifting proper path point improves the smoothing procedure. Third, continuous-curvature segments can be defined, which cross the initial path line at the position of intended turning point. Hence, there is a difference between initial path and adjusted trajectory, if these techniques were applied. Equations (18) and  (19) show that the difference is increasing if the amount of heading changes and the trajectory differs from the optimized path. Thus, there is just a small difference due to the implemented flight grid discretization. All approaches include obstacle avoidance. Finally, the resulting trajectory sections are merged to obtain the continuous trajectory.

5 Results of trajectory optimization

This section presents the results of the flight path optimization using the cost function proposed in Sect. 4.1. Overall, 12 different scenarios are tested according to the problem formulation in Sect. 2. Specifically, these are three scenarios with a freestream wind speed of ${6.5}\,\hbox {m/s}$, ${8.3}\,\hbox {m/s}$ and ${9.9}\,\hbox {m/s}$ per one of the four track directions. For each scenario, two different optimal trajectories are computed. The first one is based on the energy optimal cost function of Eq. (12). The second one is a simple shortest path optimization as described in Eq. (2). The optimization itself is in each case performed by the A-Star-Algorithm. The grid resolution of the trajectory optimization is the adjusted LES grid from the wind field prediction as described in Sect. 4.3.

Fig. 11

Comparison between shortest-way-optimization (Orange Square) and smoothed energy-optimization (Blue Square) flight track in North-South-North scenarios

Fig. 12

Comparison between shortest-way-optimization (Orange Square) and smoothed energy-optimization (Blue Square) flight track in West-East–West scenarios

Figures 11 and  12 show the results for the energy required of all scenarios. The energy savings in percentage are calculated as follows:

$$\begin{aligned} p = \frac{E_{\text {shortest way}}-E_{\text {energy opt.}}}{E_{\text {shortest way}}} \cdot 100\% \,. \end{aligned}$$

(20)

All scenarios are capable of achieving significant energy reduction. Furthermore, a clear trend of increased energy savings with higher wind speeds is present as expected. The only exception marks the East-West-${8.3}\,\hbox {m/s}$-scenario. The shortest path optimization has no knowledge about the current wind field. Thus, the same trend is more washed out and depends on head- or tailwind.

An example of the actual flight paths is given in Fig. 13, where orange is the shortest path and green the smoothed lowest energy required. Note that most of the energy savings actually come from exploiting regions of upwind and not tailwinds. This becomes clear when looking at the upwind along the trajectories, see Fig. 14. The energy optimized trajectory is leading through much larger upwind fields than the shortest way. If one observes Eq. (6), it can be seen that small differences of upwind have a greater impact on energy required than small ones of tailwinds.

Fig. 13

Flight path for $u_{W\infty } = 6.5$m/s with wind field in height of ${20}\,\hbox {m}$(→), buildings (red line), shortest way (orange line), energy optimized path after A-Star-Algorithm (dotted line), Pruning (line merged with circle) and Smoothing (green line), flying South to North

Fig. 14

Upwind for $u_{W\infty } = 6.5$m/s, South to North, flying shortest way (Orange line) and energy optimized path (Blue line)

The obtained trajectories have some characteristics worth noting for further investigations. First, all energy optimized paths are close to the rooftops of the elongated buildings. The optimizer detects the strong upwinds in front of buildings, which explains why flying alongside a building with wind perpendicular to it yields lower energy required. This effect can be called ridge lift, similar to gliding in mountains. Hence, the North–South-North paths have more savings than West-East–West paths due to more opportunities to “slide” on the roofs. Second, flying with tailwinds yields to flying higher and flying with head winds yields to flying lower. This is expected considering the wind profile in the ground’s boundary layer. This altitude behavior is seen with tailwind in Fig. 15 and headwind in Fig. 16. Still, note that the altitude changes are low. The energy required of climbing is much higher than staying on level. In contrast, descent is not able to retrieve these costs. Third, energy optimized flight paths have a remarkably longer track than the shortest way to exploit the favorable upwind locations. However, the energy savings compensate the additional way. Table 3 shows the flight tracks corresponding to all scenarios.

Fig. 15

Altitude for $u_{W\infty } = 6.5$m/s, West to East, flying shortest way (Orange line) and energy optimized path (Blue line)

Fig. 16

Altitude for $u_{W\infty } = 6.5$m/s, East to West, flying shortest way (Orange line) and energy optimized path (Blue line)

Table 3 Flight tracks

Finally, some caveats of the used methodology shall be detailed. The heuristic approach of the A-Star-Algorithm can result in inconsistencies. A-Star is not guaranteed to find an optimal solution and the heuristic function must be chosen carefully, as it (notably) influences the results. However, these drawbacks are deemed acceptable to achieve decent computation time. Further, the heuristic cost function leads to a dependence on direction. The algorithm is more likely to lead the path to a favorable upwind location if this area is close to the starting point. Each point close to the starting point has a similar high heuristic cost and very low exact costs. The costs for the detour to the upwind location are low. At the end of the path planning, close to the destination, it is better to fly directly to the destination, because the costs flying to the upwind zone first, are too expensive without knowledge of upwinds. This effect can be seen when observing that flying with headwind (East to West) requires less energy than flying with tailwind (West to East), see Fig. 12. The same can be noticed in the North–South-North track in Fig. 11 with very low head- and tailwinds. In this case, a similar energy required would be expected. Moreover, the smoothing technique used is also direction sensitive. A 90-degree change of the pre-smoothed path close to the corner of a building yields to a suggested smoothed path through the obstacle. The smoothing technique redirects the UAV in different ways depending on the starting point that influences the path modification techniques. Thus, the delivery scenarios are designed without such case, e.g., a starting point between two terraced houses and a destination perpendicular to them.

The smoothing approach can cause unfavorable paths if it misses strong upwind zones. Especially the results of the aforementioned East–West-${8.3}\,\hbox {m/s}$-scenario, differ highly from the trend of reduced energy required with increased wind speed. Hence, implementing smoothing techniques directly in the optimization process for further improvements may be ideal.

6 Conclusion

An approach to optimize 3D flight trajectories of a delivery UAV with respect to energy required was presented. The applied wind field was obtained by an LES simulation. It was shown that the tailored A-Star-Algorithm was distinguished for optimizing flight paths. The main result is that it is always beneficial to optimize the flight path at each wind speed to significantly reduce energy consumption. The test scenarios highlighted that upwinds provide the most saving potential for optimizing the flight path. In the future, more attention will be given to implementing smoothing techniques into the optimizing process to improve the presented method. Furthermore, there has some higher energy required to be respected due to short curve sections, where higher lift produces more drag. The LES wind field is planned to be validated in a series of wind tunnel experiments at TU Dresden with a scaled city model.

Appendices

Appendix A: Basic A-Star-Algorithm

The algorithm is structured in a main loop (line 1–29). At each step, it picks the node s out of the node list openPoints (line 9–11, at the beginning the starting point) with the smallest cost and processes that node. This cost is defined by summing up the exact cost of the path from the starting point to the node s, g(s), and the heuristic estimated cost from node s to the ending point described by h(s). After that, all connected neighbor nodes of node s are inserted in the node list openPoints, if they were not attended before (line 15–26). This loop of the algorithm ends the checking for a path if the ending point is reached (line 12).

Appendix B: Extended A-Star-Algorithm

At each point s, the information about the point before has to be available to compute the angle between the two connecting lines. This angle is equivalent to the heading change of the UAV. As a consequence, each point list in the algorithm has to contain the point s itself as well as the attended point before $s^{-1}$ for three reasons. First, this information is necessary for checking the heading change constraint in allocation of the neighbor points (Alg. A, line 15). Second, it is required during the reconstruction of the path at the end (Alg. A, line 13). The parent-list consists of the analyzed connection between two points and the reconstruction algorithm collects these connections and put them to one path together. It is possible that this procedure contains two paths. These can be one desired path and one path with a connection that violates the constraint. Third, the algorithm suspends points that were analyzed (Alg. A, line 28). However, this point has to be analyzed again, if the point is reached from another direction. Thus, the extended A-Star-Algorithm contains the function angleReq() that ensures the prespecified requirements of the connecting lines between the point before $s^{-1}$, current point s and following point $s'$ (line 17). Furthermore, the point lists in the basic algorithm become lists that consist of the point connections ($s \& s^{-1}$ as well as $s' \& s$) to prevent the mentioned shortcomings.

Appendix C: Basic pruning technique algorithm